A first (not well-tuned) working version of postpass scheduling for A64

1 files changed, 353 insertions, 290 deletions

diff --git a/aarch64/OpWeightsAsm.ml b/aarch64/OpWeightsAsm.ml
index c59d4695..5ba5832b 100644
--- a/aarch64/OpWeightsAsm.ml
+++ b/aarch64/OpWeightsAsm.ml
@@ -11,303 +11,366 @@
 (*                                                             *)
 (* *************************************************************)
 
-open Asmblock;;
-
 (*type called_function = (Registers.reg, AST.ident) Datatypes.sum*)
-  
-type opweights =
-  {
-    pipelined_resource_bounds : int array;
-    nr_non_pipelined_units : int;
-    latency_of_op : instruction -> int -> int;
-    resources_of_op : instruction -> int -> int array;
-    (*non_pipelined_resources_of_op : Op.operation -> int -> int array;*)
-    (*latency_of_load : AST.trapping_mode -> AST.memory_chunk -> Op.addressing -> int -> int;*)
-    (*resources_of_load : AST.trapping_mode -> AST.memory_chunk -> Op.addressing -> int -> int array;*)
-    (*resources_of_store : AST.memory_chunk -> Op.addressing -> int -> int array;*)
-    (*resources_of_cond : Op.condition -> int -> int array;*)
-    (*latency_of_call : AST.signature -> called_function -> int;*)
-    (*resources_of_call : AST.signature -> called_function -> int array;*)
-    (*resources_of_builtin : AST.external_function -> int array*)
-  };;
 
-module Cortex_A53=
-  struct
-    let resource_bounds = [| 2; 2; 1; 1 |];; (* instr ; ALU ; MAC; LSU *)
-    let nr_non_pipelined_units = 1;;
+type real_instruction =
+  | Add
+  | Adr
+  | Adrp
+  | And
+  | Asr
+  | B
+  | Bic
+  | Bl
+  | Blr
+  | Br
+  | Cbnz
+  | Cbz
+  | Cls
+  | Clz
+  | Cmn
+  | Cmp
+  | Csel
+  | Cset
+  | Eon
+  | Eor
+  | Fabs
+  | Fadd
+  | Fcmp
+  | Fcsel
+  | Fcvt
+  | Fcvtzs
+  | Fcvtzu
+  | Fdiv
+  | Fmadd
+  | Fmov
+  | Fmsub
+  | Fmul
+  | Fnmadd
+  | Fnmul
+  | Fnmsub
+  | Fneg
+  | Fsqrt
+  | Fsub
+  | Ldaxr
+  | Ldp
+  | Ldr
+  | Ldrb
+  | Ldrh
+  | Ldrsb
+  | Ldrsh
+  | Ldrsw
+  | Lr
+  | Lsl
+  | Lsr
+  | Madd
+  | Mov
+  | Movk
+  | Movn
+  | Movz
+  | Msub
+  | Nop
+  | Orn
+  | Orr
+  | Ret
+  | Rev
+  | Rev16
+  | Ror
+  | Sbfiz
+  | Sbfx
+  | Scvtf
+  | Sdiv
+  | Smulh
+  | Stlxr
+  | Stp
+  | Str
+  | Strb
+  | Strh
+  | Sub
+  | Sxtb
+  | Sxth
+  | Sxtw
+  | Tbnz
+  | Tbz
+  | Tst
+  | Ubfiz
+  | Ubfx
+  | Ucvtf
+  | Udiv
+  | Umulh
+  | Uxtb
+  | Uxth
+  | Uxtw
+  | Uxtx
+  (* Pseudo instr *)
+  | Btbl
+  | Allocframe
+  | Freeframe
+  | Builtin
+  | Cvtx2w
+
+type opweights = {
+  pipelined_resource_bounds : int array;
+  nr_non_pipelined_units : int;
+  latency_of_op : real_instruction -> int -> int;
+  resources_of_op : real_instruction -> int -> int array;
+      (*non_pipelined_resources_of_op : Op.operation -> int -> int array;*)
+      (*latency_of_load : AST.trapping_mode -> AST.memory_chunk -> Op.addressing -> int -> int;*)
+      (*resources_of_load : AST.trapping_mode -> AST.memory_chunk -> Op.addressing -> int -> int array;*)
+      (*resources_of_store : AST.memory_chunk -> Op.addressing -> int -> int array;*)
+      (*resources_of_cond : Op.condition -> int -> int array;*)
+      (*latency_of_call : AST.signature -> called_function -> int;*)
+      (*resources_of_call : AST.signature -> called_function -> int array;*)
+      (*resources_of_builtin : AST.external_function -> int array*)
+}
+
+module Cortex_A53 = struct
+  let resource_bounds = [| 2; 2; 1; 1 |]
+
+  (* instr ; ALU ; MAC; LSU *)
+  let nr_non_pipelined_units = 1
+
+  let latency_of_op (i : real_instruction) (nargs : int) =
+    match i with
+    | Add -> 1
+    | Adr -> 1
+    | Adrp -> 1 (* XXX *)
+    | And -> 1
+    | Asr -> 2
+    | B -> 1
+    | Bic -> 1
+    | Bl -> 1
+    | Blr -> 1
+    | Br -> 1
+    | Cbnz -> 1
+    | Cbz -> 1
+    | Cls -> 1
+    | Clz -> 1
+    | Cmn -> 6
+    | Cmp -> 6
+    | Csel -> 2
+    | Cset -> 2
+    | Eon -> 1
+    | Eor -> 1
+    | Fabs -> 6
+    | Fadd -> 1
+    | Fcmp -> 6
+    | Fcsel -> 1
+    | Fcvt -> 6
+    | Fcvtzs -> 6
+    | Fcvtzu -> 6
+    | Fdiv -> 50
+    | Fmadd -> 1
+    | Fmov -> 1
+    | Fmsub -> 1
+    | Fmul -> 6
+    | Fnmadd -> 1
+    | Fnmul -> 1
+    | Fnmsub -> 1
+    | Fneg -> 6
+    | Fsqrt -> 1
+    | Fsub -> 6
+    | Ldaxr -> 3
+    | Ldp -> 3
+    | Ldr -> 3
+    | Ldrb -> 3
+    | Ldrh -> 3
+    | Ldrsb -> 3
+    | Ldrsh -> 3
+    | Ldrsw -> 3
+    | Lr -> 3
+    | Lsl -> 2
+    | Lsr -> 2
+    | Madd -> 4
+    | Mov -> 1 (* XXX *)
+    | Movk -> 1
+    | Movn -> 1
+    | Movz -> 1
+    | Msub -> 4
+    | Nop -> 1
+    | Orn -> 1
+    | Orr -> 1
+    | Ret -> 1
+    | Rev -> 1
+    | Rev16 -> 1
+    | Ror -> 2
+    | Sbfiz -> 2 (* XXX *)
+    | Sbfx -> 2
+    | Scvtf -> 6
+    | Sdiv -> 50
+    | Smulh -> 4
+    | Stlxr -> 2
+    | Stp -> 2
+    | Str -> 2
+    | Strb -> 2
+    | Strh -> 2
+    | Sub -> 1
+    | Sxtb -> 1
+    | Sxth -> 1
+    | Sxtw -> 1
+    | Tbnz -> 1
+    | Tbz -> 1
+    | Tst -> 6
+    | Ubfiz -> 2 (* XXX *)
+    | Ubfx -> 2
+    | Ucvtf -> 6
+    | Udiv -> 50
+    | Umulh -> 4
+    | Uxtb -> 1
+    | Uxth -> 1
+    | Uxtw -> 1
+    | Uxtx -> 1
+    | Btbl -> 1
+    | Allocframe -> 1
+    | Freeframe -> 1
+    | Builtin -> 1
+    | Cvtx2w -> 1
+
+  let resources_of_op (i : real_instruction) (nargs : int) =
+    match i with
+    | Add -> [| 1; 1; 0; 0 |]
+    | Adr -> [| 1; 1; 0; 0 |]
+    | Adrp -> [| 1; 1; 0; 0 |]
+    | And -> [| 1; 1; 0; 0 |]
+    | Asr -> [| 1; 1; 0; 0 |]
+    | B -> [| 1; 1; 0; 0 |]
+    | Bic -> [| 1; 1; 0; 0 |]
+    | Bl -> [| 1; 1; 0; 0 |]
+    | Blr -> [| 1; 1; 0; 0 |]
+    | Br -> [| 1; 1; 0; 0 |]
+    | Cbnz -> [| 1; 1; 0; 0 |]
+    | Cbz -> [| 1; 1; 0; 0 |]
+    | Cls -> [| 1; 1; 0; 0 |]
+    | Clz -> [| 1; 1; 0; 0 |]
+    | Cmn -> [| 1; 1; 1; 0 |]
+    | Cmp -> [| 1; 1; 1; 0 |]
+    | Csel -> [| 1; 1; 1; 0 |]
+    | Cset -> [| 1; 1; 1; 0 |]
+    | Eon -> [| 1; 1; 0; 0 |]
+    | Eor -> [| 1; 1; 0; 0 |]
+    | Fabs -> [| 1; 1; 1; 0 |]
+    | Fadd -> [| 1; 1; 1; 0 |]
+    | Fcmp -> [| 1; 1; 1; 0 |]
+    | Fcsel -> [| 1; 1; 0; 0 |]
+    | Fcvt -> [| 1; 1; 1; 0 |]
+    | Fcvtzs -> [| 1; 1; 1; 0 |]
+    | Fcvtzu -> [| 1; 1; 1; 0 |]
+    | Fdiv -> [| 1; 1; 1; 0 |]
+    | Fmadd -> [| 1; 1; 0; 0 |]
+    | Fmov -> [| 1; 1; 0; 0 |]
+    | Fmsub -> [| 1; 1; 0; 0 |]
+    | Fmul -> [| 1; 1; 1; 0 |]
+    | Fnmadd -> [| 1; 1; 0; 0 |]
+    | Fnmul -> [| 1; 1; 0; 0 |]
+    | Fnmsub -> [| 1; 1; 0; 0 |]
+    | Fneg -> [| 1; 1; 1; 0 |]
+    | Fsqrt -> [| 1; 1; 0; 0 |]
+    | Fsub -> [| 1; 1; 1; 0 |]
+    | Ldaxr -> [| 1; 0; 0; 1 |]
+    | Ldp -> [| 1; 0; 0; 1 |]
+    | Ldr -> [| 1; 0; 0; 1 |]
+    | Ldrb -> [| 1; 0; 0; 1 |]
+    | Ldrh -> [| 1; 0; 0; 1 |]
+    | Ldrsb -> [| 1; 0; 0; 1 |]
+    | Ldrsh -> [| 1; 0; 0; 1 |]
+    | Ldrsw -> [| 1; 0; 0; 1 |]
+    | Lr -> [| 1; 0; 0; 1 |]
+    | Lsl -> [| 1; 1; 0; 0 |]
+    | Lsr -> [| 1; 1; 0; 0 |]
+    | Madd -> [| 1; 1; 1; 0 |]
+    | Mov -> [| 1; 1; 0; 0 |]
+    | Movk -> [| 1; 1; 0; 0 |]
+    | Movn -> [| 1; 1; 0; 0 |]
+    | Movz -> [| 1; 1; 0; 0 |]
+    | Msub -> [| 1; 1; 1; 0 |]
+    | Nop -> [| 1; 1; 0; 0 |]
+    | Orn -> [| 1; 1; 0; 0 |]
+    | Orr -> [| 1; 1; 0; 0 |]
+    | Ret -> [| 1; 1; 0; 0 |]
+    | Rev -> [| 1; 1; 0; 0 |]
+    | Rev16 -> [| 1; 1; 0; 0 |]
+    | Ror -> [| 1; 1; 0; 0 |]
+    | Sbfiz -> [| 1; 1; 0; 0 |]
+    | Sbfx -> [| 1; 1; 0; 0 |]
+    | Scvtf -> [| 1; 1; 1; 0 |]
+    | Sdiv -> [| 1; 0; 0; 0 |]
+    | Smulh -> [| 1; 1; 0; 0 |]
+    | Stlxr -> [| 1; 0; 0; 1 |]
+    | Stp -> [| 1; 0; 0; 1 |]
+    | Str -> [| 1; 0; 0; 1 |]
+    | Strb -> [| 1; 0; 0; 1 |]
+    | Strh -> [| 1; 0; 0; 1 |]
+    | Sub -> [| 1; 1; 0; 0 |]
+    | Sxtb -> [| 1; 1; 0; 0 |]
+    | Sxth -> [| 1; 1; 0; 0 |]
+    | Sxtw -> [| 1; 1; 0; 0 |]
+    | Tbnz -> [| 1; 1; 0; 0 |]
+    | Tbz -> [| 1; 1; 0; 0 |]
+    | Tst -> [| 1; 1; 1; 0 |]
+    | Ubfiz -> [| 1; 1; 0; 0 |]
+    | Ubfx -> [| 1; 1; 0; 0 |]
+    | Ucvtf -> [| 1; 1; 1; 0 |]
+    | Udiv -> [| 1; 0; 0; 0 |]
+    | Umulh -> [| 1; 1; 0; 0 |]
+    | Uxtb -> [| 1; 1; 0; 0 |]
+    | Uxth -> [| 1; 1; 0; 0 |]
+    | Uxtw -> [| 1; 1; 0; 0 |]
+    | Uxtx -> [| 1; 1; 0; 0 |]
+    | Btbl -> [| 1; 1; 0; 0 |]
+    | Allocframe -> [| 1; 0; 0; 0 |]
+    | Freeframe -> [| 1; 0; 0; 0 |]
+    | Builtin -> [| 1; 0; 0; 0 |]
+    | Cvtx2w -> [| 1; 0; 0; 0 |]
 
-    let latency_of_op (i : instruction) (nargs : int) = 1;;
-      (*match i with
-      | PArith _
-        | PArithP _ _ _
-          | Padrp _ _ -> 1 (* XXX not sure *)
-          | Pmovz _ _ _ -> 1
-          | Pmovn _ _ _ -> 1
-          | Pfmovimms _ -> 1
-          | Pfmovimmd _ -> 1
-        | PArithPP _ _ _
-          | Pmov -> 1 (* XXX not sure *)
-          | Pmovk _ _ _ -> 1
-          | Paddadr _ _ -> 1
-          | Psbfiz _ _ _ -> 2 (* XXX Maybe 1 *)
-          | Psbfx _ _ _ -> 2
-          | Pubfiz _ _ _ -> 2 (* XXX Maybe 1 *)
-          | Pubfx _ _ _ -> 2
-          | Pfmov -> 1
-          | Pfcvtds -> 6
-          | Pfcvtsd -> 6
-          | Pfabs _ -> 6
-          | Pfneg _ -> 6
-          | Pscvtf _ _ -> 6
-          | Pucvtf _ _ -> 6
-          | Pfcvtzs _ _ -> 6
-          | Pfcvtzu _ _ -> 6
-          | Paddimm _ _ -> 1
-          | Psubimm _ _ -> 1
-        | PArithPPP _ _ _ _
-          | Pasrv _ -> 2
-          | Plslv _ -> 2
-          | Plsrv _ -> 2
-          | Prorv _ -> 2 (* XXX Does not exits in Asmblockgen?? *)
-          | Psmulh -> 4
-          | Pumulh -> 4
-          | Psdiv _ -> 50
-          | Pudiv _ -> 50
-          | Paddext _ -> 2
-          | Psubext _ -> 2
-          | Pfadd _ -> 6
-          | Pfdiv _ -> 50
-          | Pfmul _ -> 6
-          | Pfsub _ -> 6
-        | PArithRR0R _ _ _ _
-          | Padd _ _ -> 2 (* XXX Maybe 6 as used in shifts *)
-          | Psub _ _ -> 2 (* XXX Maybe 6 as used in shifts *)
-          | Pand _ _ -> 1 (* XXX Maybe 2 as used in shifts *)
-          | Pbic _ _ -> 1 (* XXX Maybe 2 as used in shifts *)
-          | Peon _ _ -> 1 (* XXX Maybe 2 as used in shifts *)
-          | Peor _ _ -> 1 (* XXX Maybe 2 as used in shifts *)
-          | Porr _ _ -> 1 (* XXX Maybe 2 as used in shifts *)
-          | Porn _ _ -> 1 (* XXX Maybe 2 as used in shifts *)
-        | PArithRR0 _ _ _
-          | Pandimm _ _ -> 1
-          | Peorimm _ _ -> 1
-          | Porrimm _ _ -> 1
-        | PArithARRRR0 _ _ _ _ _
-          | Pmadd _ -> 4
-          | Pmsub _ -> 4
-        | PArithComparisonPP _ _ _ (* XXX not sure *)
-          | Pcmpext _ -> 1
-          | Pcmnext _ -> 1
-          | Pfcmp _ -> 1
-        | PArithComparisonR0R _ _ _ (* XXX not sure *)
-          | Pcmp _ _ -> 1
-          | Pcmn _ _ -> 1
-          | Ptst _ _ -> 1
-        | PArithComparisonP _ _ _ (* XXX not sure *)
-          | Pfcmp0 _ -> 1
-          | Pcmpimm _ _ -> 1
-          | Pcmnimm _ _ -> 1
-          | Ptstimm _ _ -> 1
-        | Printf.eprintf "Error NYI OpWeightsAsm"*)
+       (*let non_pipelined_resources_of_op (op : operation) (nargs : int) =
+          match op with
+          | Odiv | Odivu -> [| 29 |]
+          | Odivfs -> [| 20 |]
+          | Odivl | Odivlu | Odivf -> [| 50 |]
+          | _ -> [| -1 |];;
 
+        let resources_of_cond (cmp : condition) (nargs : int) =
+          (match cmp with
+           | Ccompf _                              (* r FP comparison *)
+             | Cnotcompf _                           (* r negation of an FP comparison *)
+             | Ccompfzero _                          (* r comparison with 0.0 *)
+             | Cnotcompfzero _                       (* r negation of comparison with 0.0 *)
+             | Ccompfs _                             (* r FP comparison *)
+             | Cnotcompfs _                          (* r negation of an FP comparison *)
+             | Ccompfszero _                         (* r equal to 0.0 *)
+             | Cnotcompfszero _                     (* r not equal to 0.0 *) ->
+              [| 1; 1; 1; 0 |]
+           | _ -> [| 1; 1; 0; 0 |] )*)
 
+  let latency_of_load trap chunk _ _ = 3
 
-    
-    let resources_of_op (i : instruction) (nargs : int) = [| 0; 0; 0; 0 |];;
-      (*match op with
-      | Omove 
-        | Ointconst _
-        | Olongconst _
-        | Ofloatconst _
-        | Osingleconst _
-        | Oaddrsymbol _
-        | Oaddrstack _
-        (* 32-bit integer arithmetic *)
-        | Oshift _
-        | Oadd 
-        | Oaddshift _
-        | Oaddimm _
-        | Oneg         
-        | Onegshift _
-        | Osub
-        | Osubshift _ -> [| 1 ; 1; 0; 0 |]
-      | Omul 
-        | Omuladd
-        | Omulsub -> [| 1; 1; 1; 0 |]
-      | Odiv 
-        | Odivu -> [| 1; 0; 0; 0 |]
-      | Oand 
-        | Oandshift _
-        | Oandimm _
-        | Oor
-        | Oorshift _
-        | Oorimm _
-        | Oxor
-        | Oxorshift _
-        | Oxorimm _
-        | Onot
-        | Onotshift _
-        | Obic
-        | Obicshift _
-        | Oorn
-        | Oornshift _
-        | Oeqv
-        | Oeqvshift _
-        | Oshl
-        | Oshr
-        | Oshru
-        | Oshrximm _
-        | Ozext _
-        | Osext  _
-        | Oshlzext _
-        | Oshlsext _
-        | Ozextshr _
-        | Osextshr _
-  
-(* 64-bit integer arithmetic *)
-        | Oshiftl _
-        | Oextend _
-        | Omakelong
-        | Olowlong 
-        | Ohighlong
-        | Oaddl 
-        | Oaddlshift _
-        | Oaddlext _
-        | Oaddlimm _
-        | Onegl        
-        | Oneglshift _
-        | Osubl
-        | Osublshift _
-        | Osublext _ -> [| 1 ; 1 ; 0; 0 |]
-      | Omull 
-        | Omulladd 
-        | Omullsub
-        | Omullhs 
-        | Omullhu -> [| 1 ; 1 ; 1; 0 |]
-      | Odivl 
-        | Odivlu -> [| 1; 0; 0; 0 |]
-      | Oandl 
-        | Oandlshift _
-        | Oandlimm _
-        | Oorl
-        | Oorlshift _
-        | Oorlimm _
-        | Oxorl
-        | Oxorlshift _
-        | Oxorlimm _
-        | Onotl
-        | Onotlshift _
-        | Obicl
-        | Obiclshift _
-        | Oornl
-        | Oornlshift _
-        | Oeqvl
-        | Oeqvlshift _
-        | Oshll
-        | Oshrl
-        | Oshrlu 
-        | Oshrlximm _
-        | Ozextl _
-        | Osextl _
-        | Oshllzext _
-        | Oshllsext _
-        | Ozextshrl _
-        | Osextshrl _ -> [| 1; 1; 0; 0 |]
-      (* 64-bit floating-point arithmetic *)
-      | Onegf                                               (* r [rd = - r1] *)
-        | Oabsf                                               (* r [rd = abs(r1)] *)
-        | Oaddf                                               (* r [rd = r1 + r2] *)
-        | Osubf                                               (* r [rd = r1 - r2] *)
-        | Omulf                                               (* r [rd = r1 * r2] *)
-        | Odivf
-        (* 32-bit floating-point arithmetic *)
-        | Onegfs                                              (* r [rd = - r1] *)
-        | Oabsfs                                              (* r [rd = abs(r1)] *)
-        | Oaddfs                                              (* r [rd = r1 + r2] *)
-        | Osubfs                                              (* r [rd = r1 - r2] *)
-        | Omulfs                                              (* r [rd = r1 * r2] *)
-        | Odivfs                                              (* r [rd = r1 / r2] *)
-        | Osingleoffloat                                      (* r [rd] is [r1] truncated to single-precision float *)
-        | Ofloatofsingle                                      (* r [rd] is [r1] extended to double-precision float *)
-(*  Conversions between int and float *)
-        | Ointoffloat                                         (* r [rd = signed_int_of_float64(r1)] *)
-        | Ointuoffloat                                        (* r [rd = unsigned_int_of_float64(r1)] *)
-        | Ofloatofint                                         (* r [rd = float64_of_signed_int(r1)] *)
-        | Ofloatofintu                                        (* r [rd = float64_of_unsigned_int(r1)] *)
-        | Ointofsingle                                        (* r [rd = signed_int_of_float32(r1)] *)
-        | Ointuofsingle                                       (* r [rd = unsigned_int_of_float32(r1)] *)
-        | Osingleofint                                        (* r [rd = float32_of_signed_int(r1)] *)
-        | Osingleofintu                                       (* r [rd = float32_of_unsigned_int(r1)] *)
-        | Olongoffloat                                        (* r [rd = signed_long_of_float64(r1)] *)
-        | Olonguoffloat                                       (* r [rd = unsigned_long_of_float64(r1)] *)
-        | Ofloatoflong                                        (* r [rd = float64_of_signed_long(r1)] *)
-        | Ofloatoflongu                                       (* r [rd = float64_of_unsigned_long(r1)] *)
-        | Olongofsingle                                       (* r [rd = signed_long_of_float32(r1)] *)
-        | Olonguofsingle                                      (* r [rd = unsigned_long_of_float32(r1)] *)
-        | Osingleoflong                                       (* r [rd = float32_of_signed_long(r1)] *)
-        | Osingleoflongu                                      (* r [rd = float32_of_unsigned_int(r1)] *)
-        -> [| 1 ; 1; 1; 0 |]
-         
-(* Boolean tests *)
-      | Ocmp cmp | Osel (cmp, _) ->
-         (match cmp with
-          | Ccompf _                              (* r FP comparison *)
-            | Cnotcompf _                           (* r negation of an FP comparison *)
-            | Ccompfzero _                          (* r comparison with 0.0 *)
-            | Cnotcompfzero _                       (* r negation of comparison with 0.0 *)
-            | Ccompfs _                             (* r FP comparison *)
-            | Cnotcompfs _                          (* r negation of an FP comparison *)
-            | Ccompfszero _                         (* r equal to 0.0 *)
-            | Cnotcompfszero _                     (* r not equal to 0.0 *) ->
-             [| 1; 1; 1; 0 |]
-          | _ -> [| 1; 1; 0; 0 |] );;
-    
-    let non_pipelined_resources_of_op (op : operation) (nargs : int) =
-      match op with
-      | Odiv | Odivu -> [| 29 |]
-      | Odivfs -> [| 20 |]
-      | Odivl | Odivlu | Odivf -> [| 50 |]
-      | _ -> [| -1 |];;
-    
-    let resources_of_cond (cmp : condition) (nargs : int) =
-      (match cmp with
-       | Ccompf _                              (* r FP comparison *)
-         | Cnotcompf _                           (* r negation of an FP comparison *)
-         | Ccompfzero _                          (* r comparison with 0.0 *)
-         | Cnotcompfzero _                       (* r negation of comparison with 0.0 *)
-         | Ccompfs _                             (* r FP comparison *)
-         | Cnotcompfs _                          (* r negation of an FP comparison *)
-         | Ccompfszero _                         (* r equal to 0.0 *)
-         | Cnotcompfszero _                     (* r not equal to 0.0 *) ->
-          [| 1; 1; 1; 0 |]
-       | _ -> [| 1; 1; 0; 0 |] )*);;
-    
-    let latency_of_load trap chunk _ _ = 3;;
-    let latency_of_call _ _ = 6;;
-    
-    let resources_of_load trap chunk addressing nargs = [| 1; 0; 0; 1 |];;
-    
-    let resources_of_store chunk addressing nargs = [| 1; 0; 0; 1 |];;
-    
-    let resources_of_call _ _ = resource_bounds;;
-    let resources_of_builtin _ = resource_bounds;;
-  end;;
+  let latency_of_call _ _ = 6
+
+  let resources_of_load trap chunk addressing nargs = [| 1; 0; 0; 1 |]
+
+  let resources_of_store chunk addressing nargs = [| 1; 0; 0; 1 |]
+
+  let resources_of_call _ _ = resource_bounds
+
+  let resources_of_builtin _ = resource_bounds
+end
 
 let get_opweights () : opweights =
   (*match !Clflags.option_mtune with*)
   (*| "cortex-a53" | "cortex-a35" | "" ->*)
-     {
-       pipelined_resource_bounds = Cortex_A53.resource_bounds;
-       nr_non_pipelined_units = Cortex_A53.nr_non_pipelined_units;
-       latency_of_op = Cortex_A53.latency_of_op;
-       resources_of_op = Cortex_A53.resources_of_op;
-       (*non_pipelined_resources_of_op = Cortex_A53.non_pipelined_resources_of_op;*)
-       (*latency_of_load = Cortex_A53.latency_of_load;*)
-       (*resources_of_load = Cortex_A53.resources_of_load;*)
-       (*resources_of_store = Cortex_A53.resources_of_store;*)
-       (*resources_of_cond = Cortex_A53.resources_of_cond;*)
-       (*latency_of_call = Cortex_A53.latency_of_call;*)
-       (*resources_of_call = Cortex_A53.resources_of_call;*)
-       (*resources_of_builtin = Cortex_A53.resources_of_builtin*)
-     }
-  (*| xxx -> failwith (Printf.sprintf "unknown -mtune: %s" xxx);;*)
+  {
+    pipelined_resource_bounds = Cortex_A53.resource_bounds;
+    nr_non_pipelined_units = Cortex_A53.nr_non_pipelined_units;
+    latency_of_op = Cortex_A53.latency_of_op;
+    resources_of_op = Cortex_A53.resources_of_op
+      (*non_pipelined_resources_of_op = Cortex_A53.non_pipelined_resources_of_op;*)
+      (*latency_of_load = Cortex_A53.latency_of_load;*)
+      (*resources_of_load = Cortex_A53.resources_of_load;*)
+      (*resources_of_store = Cortex_A53.resources_of_store;*)
+      (*resources_of_cond = Cortex_A53.resources_of_cond;*)
+      (*latency_of_call = Cortex_A53.latency_of_call;*)
+      (*resources_of_call = Cortex_A53.resources_of_call;*)
+      (*resources_of_builtin = Cortex_A53.resources_of_builtin*);
+  }
+
+(*| xxx -> failwith (Printf.sprintf "unknown -mtune: %s" xxx);;*)